FLUX 3 Max Is Image-to-Video. Start From a Locked Still.

fal launched FLUX 3 Max on 19 October 2026 as a post-trained video model from fal Research and Black Forest Labs. The product page, updated 18 October 2026, lists two endpoints at launch: text-to-video and image-to-video. fal says a 10-second 720p clip generates in about 12 seconds. If you already have a mug, bottle, jacket, or ad hero that matches the unit you ship, do not describe it from memory into text-to-video. Lock the still first. Name what stays. Ask for one motion. Inspect. Overlay type last. FreeGPTImg is a GPT Image 2 workspace: 20 daily free credits, no credit card required. It does not run FLUX 3 Max and it does not export video.

Keep the still loop in FreeGPTImg. Stay on GPT Image 2 when that is the label on the screen, invent a first frame only when you have no photo in text to image or the free AI image generator, edit a locked hero in image to image, keep materials honest in AI product photography, place one in-use scene in the lifestyle product photo tool, and recrop a feed frame in AI social media ads.

What launched this week, and what did not

Three names get mixed in one folder. Mixing them is how a fast video model turns a honest catalog JPEG into a second spout.

1. FLUX 3 (Black Forest Labs). fal’s own comparison says this is Black Forest Labs’ release, with draft, keyframe, and extend endpoints. It is the base model. It is not the post-trained Max route.
2. FLUX 3 Max (fal Research). fal Research post-trained FLUX 3 with Black Forest Labs. At launch it is text-to-video and image-to-video. fal cites Artificial Analysis’s preliminary image-to-video arena: 1,139 Elo, tied for first, 50 points above FLUX 3 (1,089), and about 10× faster than FLUX 3 on a 10-second clip (12.45 s vs 126.44 s). The same page compares 12.45 s against 277.80 s for Seedance 2.5 at the same 1,139 Elo. Treat those numbers as fal’s published comparison, not as a ranking badge you should bake into pixels.
3. A GPT Image 2 still. That is the first frame you own. FreeGPTImg generates and edits stills here. It is not a FLUX playground.

Pricing on fal’s page: 75% off until 7 November 2026, $0.0425 per second of 720p output, list $0.17, billed per second, no subscription. Recheck fal’s pricing page before you quote a teammate. Product Hunt listed the launch on 23 October 2026. None of those dates belong in the still.

The cheap failure this week is not “I forgot the API key.” It is “text-to-video sounded faster, so I typed the product and got a cousin SKU that now moves.”

Name the job: still, then motion

Weak sets fail because they ask one prompt to be a launch announcement, a new white MAIN, a 9:16 story, and a “#1 I2V” badge. Name the job first.

Job	Start from	Generate in	Keep	Wrong move
Locked catalog still already honest	That JPEG	Image-to-video later	Silhouette, label, pack count	Text-to-video from a new paragraph
One new view of the same SKU	Locked hero	Image-to-image, then I2V	Lighting family, materials	“Make it more cinematic and more Max”
Lifestyle field for a 10-second clip	Locked hero	Image-to-image, field only	SKU, printed marks	Inventing a second colorway that then moves
Social first frame	Locked hero	Recrop after the still is honest	Empty headline	Stretching a cousin SKU into 9:16
New SKU, no photo yet	A real product photo if you have one	Text-to-image, then lock	Empty type	Baking “FLUX 3 Max” into pixels
Motion pass	Locked still	FLUX 3 Max image-to-video, elsewhere	Same SKU, one motion	Asking the video model to redesign the handle
Speed or price note	Don’t generate it	Editor, overlay last	Empty Elo / $/sec	“#1 I2V 12s” badge on the product

A text-to-video pass invents. An image-to-video pass should borrow. If you already paid for a still that matches the unit you ship, that file is truth. Speed does not get to redesign the handle.

Still checklist before FLUX 3 Max

Use this sequence on GPT Image 2, then take the file to whatever image-to-video tab you actually use. Do not paste a launch blog into a creator prompt and expect the model to keep the SKU.

1. Find the locked hero. One JPEG or PNG that already matches the unit you ship. Even light, product filling most of the frame, quiet field. If you do not have that file, stop and shoot or scan one. Do not invent it from a paragraph because video got cheaper this week.
2. Write keep / change / do-not in five lines. Keep silhouette, materials, pack count, printed logo if it exists on the object, and the exact SKU. For the still, change only one noun: crop, view, field, or light. For the later video pass, change only one motion: a slow push-in, a gentle rotation, steam, fabric, or a hand entering once. Do not add a second logo, a wax seal, a QR code, or a cousin SKU.
3. Name the engine on the screen. If the tab says GPT Image 2, write gpt-image-2 into the still filename. If you later run FLUX 3 Max, write that on the video file, not on the JPEG. Do not relabel a GPT Image 2 still as Max.
4. Generate or edit the still first. Image-to-image starts from pixels. Text-to-image is only for a first frame when no photo exists. A prompt without the file is invention wearing a costume.
5. Ask for empty type. Prices, percents, “FLUX 3 Max,” “12s,” “NEW,” Elo scores, and marketplace badges never belong in the generator if they must be exact.
6. Leave motion room. Product in the middle third. Empty lower band if a later overlay needs it. Do not fill every edge with props that will swim when the camera moves.
7. Generate one still. One pass. File names like mug-three-quarter-gpt-image-2.jpg.
8. Inspect the subject before the story. A sixth finger, a new spout, a badge reading “Max,” or a tag that says “30% OFF” are stop conditions. Hold the still next to the real sample. If a stranger would think this unit has a colorway you do not ship, discard it. Motion will not fix a dishonest still. It will advertise it.
9. Only then image-to-video. Upload the locked still. Repeat the keep list. Name one motion. Ten seconds is enough. Do not also ask for a new pack, a new location, and a talking presenter.
10. Overlay copy last. In an editor, set the real headline and only the claims a manager signed. Recrop to 1:1, 4:5, or 9:16 after the subject is honest. Do not ask either model to spell the offer.

Prompt formula for the still

GPT Image 2 still for a later image-to-video pass, not a FLUX 3 Max demo and not a ranking poster.
Start from the uploaded photo of one [SKU], or describe one SKU only if no photo exists.
Keep: exact silhouette, materials, pack count, printed marks, lighting family.
Change only: [view / field / crop / light].
Leave empty lower third for later overlay. Product in the middle third.
Text in image: none. No “FLUX 3 Max”, “12s”, “#1”, “NEW”, prices, watermarks, or marketplace badges.
Do not add: extra SKUs, extra colorways, extra accessories, celebrity hands, children, camera chrome.

Prompt examples

Locked mug, three-quarter view. “Image-to-image: keep this exact ceramic mug silhouette, glaze, handle curve, and studio lighting. Change only to a three-quarter view. Quiet field, empty lower band. No extra objects, no text, no second mug.”

White field, then motion later. “Image-to-image: keep this exact bottle silhouette, label, and cap. Change only the field to pure white, even catalog light. No extra bottles, no ‘FLUX 3 Max’ type, no droplets that were not on the sample.”

Lifestyle first frame for a 10-second clip. “Image-to-image: keep this exact kettle silhouette, steel finish, and handle. Change only the field to a quiet oak counter, warm side light, empty lower band. No extra kettles, no recipe card text.”

9:16 social still, type empty. “4:5 then 9:16 recrop of the locked hero, product in the middle third, empty lower band for later overlay. No ‘30% off’ type, no Elo scores, no fal chrome.”

Motion brief you take elsewhere, after the still is honest. “Image-to-video from this exact still. Keep the SKU, label, and lighting. Change only: slow 10-second push-in, no new objects, no on-screen text.”

Do not generate this. “Infographic that says ‘FLUX 3 Max #1 I2V, 12.45 seconds, 75% off, guaranteed ad approval,’ with a fake arena badge and a cousin SKU that now rotates.”

Review the pixels before you animate them

Zoom to 100%. A phone grid hides a rewritten logo. A 10-second clip will not hide it either.

• Subject honesty. Same silhouette, materials, finish, pack count, and SKU. If a stranger would think this unit has a feature it does not ship, discard the still.
• Edit honesty. Only the noun you named should move in the still. If the handle, spout, stitch line, or label type also moved, the pass failed.
• Motion honesty. One motion. Steam is not a new lid. A slow orbit is not a second colorway. If the video invents a logo, stop; the still was the last honest file.
• Model honesty. Write the model you used into the file name. A GPT Image 2 still is GPT Image 2. A FLUX 3 Max clip is that clip. Do not stamp Max on the JPEG.
• Empty type. Prices, percents, model IDs, Elo, and “NEW” stay empty of anything you need people to read.
• People and file. No recognizable generated faces you cannot clear. No children. JPEG or PNG, RGB, enough pixels for a later 720p clip, no watermark, no QR. Preview on a phone and as a vertical crop.
• Provenance. A video launch is not a license. If a platform requires AI disclosure, follow that policy. Recheck fal’s product page before you tell a teammate the endpoints, the price, or the 12-second claim.

Limitations you should not skip

A generated still is not ad-network approval, not a commercial-rights grant, and not proof that FLUX 3 Max ran. FreeGPTImg does not call FLUX 3 Max, does not bill per second of video, and does not turn a JPEG into a 10-second clip. The default loop here is GPT Image 2 stills. Do not generate medical seals, “#1 2026,” or sale badges unless those claims are already on the real label and a manager signed them.

fal’s page is explicit about what Max is not: it is not FLUX 3’s draft / keyframe / extend set. Endpoints at launch are text-to-video and image-to-video, information dated 18 October 2026. fal is still post-training the model. Speed and Elo figures are fal’s published comparison against FLUX 3 and Seedance 2.5 on a preliminary board. Do not treat them as a guaranteed result on your SKU.

Models invent letters when you mention a launch, a percent, or a model ID as scene dressing. Shorten the brief. One SKU. One still. One motion. Empty type. A new paragraph is not a backup of the photo you already had. If the locked hero is honest, animate that. If it is not honest, reshoot. Do not ask a fast video model to remember a product you never locked.

FAQ

What is FLUX 3 Max, and when did it launch?

fal announced FLUX 3 Max on 19 October 2026. It is a post-trained video model from fal Research, built with Black Forest Labs on FLUX 3. fal’s product page lists text-to-video and image-to-video at launch, a 10-second 720p clip in about 12 seconds, and a 75% discount until 7 November 2026 at $0.0425 per second of 720p output.

Should I text-to-video my product from a paragraph?

No. If you already have a locked photo that matches the unit you ship, use that file as the image-to-video input. Text-to-video is for a first clip when no honest still exists. Image-to-video is for motion on a SKU you already trust.

Does FreeGPTImg generate FLUX 3 Max video?

No. FreeGPTImg is a GPT Image 2 workspace: 20 daily free credits, no credit card required. Use it to lock the still. Run image-to-video on the platform that actually hosts FLUX 3 Max.

Is FLUX 3 Max the same as FLUX 3?

No. fal says FLUX 3 is Black Forest Labs’ release, with draft, keyframe, and extend endpoints. FLUX 3 Max is the post-trained version focused on text-to-video and image-to-video.

Why did the clip add a second handle or the words “Max”?

The prompt mentioned a launch, a ranking, or “make it more cinematic” as scene dressing. Shorten it. Upload the locked still. Name one motion. Empty type. Generate again.

Can I put prices or “#1 I2V” in the first frame?

Not in the generator, not if those words must be exact. Leave type empty. Overlay the real offer in an editor after the still and the motion both match the product you ship.

Sources

• fal: FLUX 3 Max
• PR Newswire: fal Launches FLUX 3 Max
• Product Hunt: FLUX 3 Max launch
• Black Forest Labs: bfl.ai
• OpenAI: Usage policies

Lock the still on FreeGPTImg first

Pick the JPEG that already matches the unit you ship. Write five lines: keep, one still change, one later motion, empty type. Open FreeGPTImg, stay on GPT Image 2 when that is the label on the screen, invent a first frame only if you have no photo in text to image or the free AI image generator, run the edit in image to image, keep materials honest in AI product photography, place one in-use scene with the lifestyle product photo tool, recrop a feed frame in AI social media ads, and only then take the locked still into image-to-video.